Human behavior recognition from sensor observations is an important topic in both artificial intelligence and mobile computing. It is also a difficult task as the sensor and behavior data are usually noisy and limited. In this thesis, we first introduce the three major problems in human behavior recognition, including location estimation, activity recognition and mobile recommendation. Solving these three problems helps to answer the typical questions in human behavior recognition, such as where a user is, what s/he is doing and whether s/he will be interested in doing something at somewhere. In our attempt to solve these problems, we find that in practice the biggest challenge comes from the data sparsity. Such data sparsity can be because we have limited labeled data for new contexts...[ Read more ]

Human behavior recognition from sensor observations is an important topic in both artificial intelligence and mobile computing. It is also a difficult task as the sensor and behavior data are usually noisy and limited. In this thesis, we first introduce the three major problems in human behavior recognition, including location estimation, activity recognition and mobile recommendation. Solving these three problems helps to answer the typical questions in human behavior recognition, such as where a user is, what s/he is doing and whether s/he will be interested in doing something at somewhere. In our attempt to solve these problems, we find that in practice the biggest challenge comes from the data sparsity. Such data sparsity can be because we have limited labeled data for new contexts in localization, or limited sensor data for users / activities in activity recognition, or limited activity data for mobile recommendation. In order to address these challenges, we propose learning methods which can effectively incorporate domain-dependent auxiliary data in training and thus greatly relieve the sparsity problem. We conduct empirical studies with real-world data sets, and demonstrate the effectiveness of our algorithms over the competing baselines.